Leaf Occlusion Research Articles

Object Detection Algorithm for Citrus Fruits Based on Improved YOLOv5 Model

To address the challenges of missed and false detections in citrus fruit detection caused by environmental factors such as leaf occlusion, fruit overlap, and variations in natural light in hilly and mountainous orchards, this paper proposes a citrus detection model based on an improved YOLOv5 algorithm. By introducing receptive field convolutions with full 3D weights (RFCF), the model overcomes the issue of parameter sharing in convolution operations, enhancing detection accuracy. A focused linear attention (FLA) module is incorporated to improve the expressive power of the self-attention mechanism while maintaining computational efficiency. Additionally, anchor boxes were re-clustered based on the shape characteristics of target objects, and the boundary box loss function was improved to Foal-EIoU, boosting the model’s localization ability. Experiments conducted on a citrus fruit dataset labeled using LabelImg, collected from hilly and mountainous areas, showed a detection precision of 95.83% and a mean average precision (mAP) of 79.68%. This research not only significantly improves detection performance in complex environments but also provides crucial data support for precision tasks such as orchard localization and intelligent picking, demonstrating strong potential for practical applications in smart agriculture.

Identification of weeds in cotton fields at various growth stages using color feature techniques

Weeds affect the growth and yield of cotton (Gossypium hirsutum), which is the most prevalent commercially viable non-food crop. Accurate differentiation between weeds and cotton plants is crucial for implementing precise weed-control strategies, optimizing crop yields, and minimizing herbicide usage. In this study, we aimed to investigate the effectiveness of using color feature techniques combined with Otsu's method (OTSU) to distinguish weeds from cotton plants at various growth stages. The results revealed that the image segmentation approach achieved a recognition rate of 74.17% during the second growth stage. In the third growth stage, the weed identification process primarily targeted lambsquarters (Chenopodium album). The standard deviation (SD) difference between the red (R), green (G), and blue (B) components of the plant images revealed that the difference between the SD values of R and B was <5, which could serve as an effective threshold for identifying lambsquarters. Additionally, the recognition rates for cotton and lambsquarters were 71.40% and 92.90%, respectively, and the overall recognition rate reached 82.10% during the identification process. The results of this study have practical implications for weed management practices, such as targeted herbicide application, mitigation of environmental impacts, and contribution to higher crop yields. Future research will focus on refining these methods by exploring advanced image processing techniques that integrate deep learning for automatic feature extraction, which will be evaluated for handling complex agronomic scenarios, such as leaf occlusion, varying growth stages, and image noise.

Citrus recognition in orchard scene based on modified HSV-morphology method

The orchard has always been an important scene for citrus pickers, and the existence of factors such as leaf occlusion and color similarity always lead to the difficulty of robot recognition. Based on a citrus image dataset collected from actual orchard scenes, we extract image features and build a mathematical model to identify and count the number of oranges in each image, and show the distribution of apples in the entire dataset. Firstly, this paper establishes a model based on HSV method, focusing on recognizing yellow citrus. Secondly, with the further analysis of the data set, it is found that the cyan citrus exists. Therefore, this paper defines the empirical HSV range of two kinds of citrus at the same time and introduces contour processing to optimize the model. In order to accurately distinguish the immature citrus from the leaves with similar colors and the citrus occluded by leaves, this paper uses more detailed color threshold setting and morphological operation, and makes the eccentricity of the contour less than 0.85. At the same time, considering the size and overlap of citrus, morphological operations such as dilation and erosion and contour detection techniques are applied to further improve the recognition ability of the model. Finally, the experimental results show that the average accuracy of the established model reaches 92.1%, which has a good recognition effect on citrus.

A deep learning-based method for estimating the main stem length of sweet potato seedlings

The length of sweet potato seedling main stems is critical for harvest decisions, but accurately measuring it is challenging due to leaf occlusion and self-curvature. This study proposes a method for stem reconstruction and length estimation using a two-stage classifier, feature extraction, and semantic segmentation to obtain stem masks. Image processing separates clustered skeletons into singular lines for detailed analysis. Object detection identifies roots, stem-leaf intersections, and establishes a search relationship among segments. Stems are reconstructed based on directional, distance, and curvature constraints, then divided into four parts for length estimation using stereo depth information. Experimental results demonstrate effectiveness under various conditions: un-occluded, partially occluded, curved stems, and low-density clusters. Median length estimation error is 1.7 cm, with maximum and minimum errors of 3.2 cm and 0.3 cm respectively. This automated method provides a reliable solution for sweet potato seedling harvest and grading.

Research on High Performance Apple Recognition Model Construction Strategy Based on Deep Learning

In order to improve the working efficiency and environmental adaptability of apple picking robots, which can realize accurate recognition and localization, ripening assessment, and quality assessment under complex situations such as "leaf occlusion", "branch occlusion", "fruit occlusion", "mixed occlusion", etc., this paper proposes a Faster R-CNN-based apple recognition method. "mixed occlusion" and other complex situations to achieve accurate recognition and localization, maturity assessment, quality assessment, the paper proposes an apple recognition method based on Faster R-CNN. The method combines the cv2.inRange and cv2.fitEllipse functions in OpenCV to optimize the model, and performs color segmentation and shape fitting by setting up color and shape thresholds to recognize apples. The trained model was tested to have a recall of 88.44% under the validation set, an accuracy of 90.35%, an F1 value of 89.38%, and a recognition time of about 0.3 s per image.The experimental results showed that the number of apples in the dataset showed a normal distribution, with the number distributed more in the range of 0 to 10. Candidate frames are screened by RPN, and apple coordinates are derived based on the coordinate transformation formula, and the experimental results show that the apples are more densely distributed near the range of x=100, y=100. The image color distribution was extracted through color histogram, and the entropy value of GLCM, contrast as the texture feature of apples, and the classification prediction of apple ripeness was carried out by using Random Forest Model, and the accuracy of the model prediction was 83.2%, the precision rate was 84.1%, and the recall rate was 83.5%, and the experimental results showed that 88.7% of the apples were ripe. Image edges were extracted by the Canny edge detection algorithm and the area values of pixel points within the edges were calculated. The ratio of pixel area converted to apple mass was estimated to be 0.007 based on the image resolution and apple density, and the test results indicated that the mass of apples was mostly concentrated in the range of 0~100g.

RSR-YOLO: a real-time method for small target tomato detection based on improved YOLOv8 network

In tomato producing fields, automated large-area detection method is critical for fruit health monitoring and harvesting. However, due to the limited feature information included in tomatoes, large-area detection across long distances results in more missing or incorrect detections. To address this issue, this research proposes an improved YOLOv8 network, RSR-YOLO, for long-distance identification of tomato fruits. Firstly, this paper designs a partial group convolution (PgConv) and furthermore an innovative FasterNet (IFN) module for feature extraction, taking into account the impact of split operations on the computational complexity of the backbone network. The IFN module is lightweight and efficient, which improves the detection accuracy and real-time detection performance of the model. Secondly, this research combines the Gather and Distribute mechanism (GD) and redesigns the feature fusion module to implement the extraction and fusion of various levels of tomato features, given the critical significance that low-dimensional features play in small target recognition and localization. Finally, Repulsion Loss is used in this paper to examine the impact of fruit overlap and leaf occlusion on detection outcomes. RSR-YOLO achieves precision, recall, F1 score, and mean average precision (mAP@0.5) of 91.6%, 85.9%, 88.7%, and 90.7%, respectively, marking increases of 4.2%, 4%, 4.2%, and 3.6% compared to YOLOv8n. In addition, this paper designs a specialized Graphical User Interface (GUI) for the real-time detection task of tomatoes.

基于DeepLabV3的复杂果园环境下樱桃分割与识别

Aiming at the problems of less research on cherry segmentation and identification, with slow recognition speed and low classification accuracy in agricultural products, a method based on DeepLabV3 was proposed to realize the rapid segmentation and identification of cherry in complex orchard environment. Complex environment mainly includes front lighting, back lighting, cloudy and rainy days, single fruit, multi fruit, fruit overlap, and branch and leaf occlusion. This model proposed the Atrous Spatial Pyramid Pooling (ASPP) module to effectively extract multi-scale contextual information, and solved the problem of target segmentation at multiple scales. The obtained data was divided into training, validation and testing sets in 7:1:2 ratios, and the residual network 50 (ResNet50) was selected as backbone of the DeepLabV3. Experimental results show that the algorithm in this paper can segment cherry quickly and accurately, the mean intersection over union (MIoU) was 91.06%, the mean pixel accuracy (MPA) was 93.05%, and the kappa coefficient was 0.89, which was better than fully convolutional networks (FCN), SegNet, DeepLabV1 and DeepLabV2. It is demonstrated that this study can provide technical support for intelligent segmentation of agricultural products.

A method for small-sized wheat seedlings detection: from annotation mode to model construction

The number of seedlings is an important indicator that reflects the size of the wheat population during the seedling stage. Researchers increasingly use deep learning to detect and count wheat seedlings from unmanned aerial vehicle (UAV) images. However, due to the small size and diverse postures of wheat seedlings, it can be challenging to estimate their numbers accurately during the seedling stage. In most related works in wheat seedling detection, they label the whole plant, often resulting in a higher proportion of soil background within the annotated bounding boxes. This imbalance between wheat seedlings and soil background in the annotated bounding boxes decreases the detection performance. This study proposes a wheat seedling detection method based on a local annotation instead of a global annotation. Moreover, the detection model is also improved by replacing convolutional and pooling layers with the Space-to-depth Conv module and adding a micro-scale detection layer in the YOLOv5 head network to better extract small-scale features in these small annotation boxes. The optimization of the detection model can reduce the number of error detections caused by leaf occlusion between wheat seedlings and the small size of wheat seedlings. The results show that the proposed method achieves a detection accuracy of 90.1%, outperforming other state-of-the-art detection methods. The proposed method provides a reference for future wheat seedling detection and yield prediction.

Maize plant detection using UAV-based RGB imaging and YOLOv5.

In recent years, computer vision (CV) has made enormous progress and is providing great possibilities in analyzing images for object detection, especially with the application of machine learning (ML). Unmanned Aerial Vehicle (UAV) based high-resolution images allow to apply CV and ML methods for the detection of plants or their organs of interest. Thus, this study presents a practical workflow based on the You Only Look Once version 5 (YOLOv5) and UAV images to detect maize plants for counting their numbers in contrasting development stages, including the application of a semi-auto-labeling method based on the Segment Anything Model (SAM) to reduce the burden of labeling. Results showed that the trained model achieved a mean average precision (mAP@0.5) of 0.828 and 0.863 for the 3-leaf stage and 7-leaf stage, respectively. YOLOv5 achieved the best performance under the conditions of overgrown weeds, leaf occlusion, and blurry images, suggesting that YOLOv5 plays a practical role in obtaining excellent performance under realistic field conditions. Furthermore, introducing image-rotation augmentation and low noise weight enhanced model accuracy, with an increase of 0.024 and 0.016 mAP@0.5, respectively, compared to the original model of the 3-leaf stage. This work provides a practical reference for applying lightweight ML and deep learning methods to UAV images for automated object detection and characterization of plant growth under realistic environments.

Improved Faster Region-Based Convolutional Neural Networks (R-CNN) Model Based on Split Attention for the Detection of Safflower Filaments in Natural Environments

The accurate acquisition of safflower filament information is the prerequisite for robotic picking operations. To detect safflower filaments accurately in different illumination, branch and leaf occlusion, and weather conditions, an improved Faster R-CNN model for filaments was proposed. Due to the characteristics of safflower filaments being dense and small in the safflower images, the model selected ResNeSt-101 with residual network structure as the backbone feature extraction network to enhance the expressive power of extracted features. Then, using Region of Interest (ROI) Align improved ROI Pooling to reduce the feature errors caused by double quantization. In addition, employing the partitioning around medoids (PAM) clustering was chosen to optimize the scale and number of initial anchors of the network to improve the detection accuracy of small-sized safflower filaments. The test results showed that the mean Average Precision (mAP) of the improved Faster R-CNN reached 91.49%. Comparing with Faster R-CNN, YOLOv3, YOLOv4, YOLOv5, and YOLOv6, the improved Faster R-CNN increased the mAP by 9.52%, 2.49%, 5.95%, 3.56%, and 1.47%, respectively. The mAP of safflower filaments detection was higher than 91% on a sunny, cloudy, and overcast day, in sunlight, backlight, branch and leaf occlusion, and dense occlusion. The improved Faster R-CNN can accurately realize the detection of safflower filaments in natural environments. It can provide technical support for the recognition of small-sized crops.

Prototype Network for Predicting Occluded Picking Position Based on Lychee Phenotypic Features

The automated harvesting of clustered fruits relies on fast and accurate visual perception. However, the obscured stem diameters via leaf occlusion lack any discernible texture patterns. Nevertheless, our human visual system can often judge the position of harvesting points. Inspired by this, the aim of this paper is to address this issue by leveraging the morphology and the distribution of fruit contour gradient directions. Firstly, this paper proposes the calculation of fruit normal vectors using edge computation and gradient direction distribution. The research results demonstrate a significant mathematical relationship between the contour edge gradient and its inclination angle, but the experiments show that the standard error projected onto the Y-axis is smaller, which is evidently more conducive to distinguishing the gradient distribution. Secondly, for the front view of occluded lychee clusters, a fully convolutional, feature prototype-based one-stage instance segmentation network is proposed, named the lychee picking point prediction network (LP3Net). This network can achieve high accuracy and real-time instance segmentation, as well as for occluded and overlapping fruits. Finally, the experimental results show that the LP3Net based on this study, along with lychee phenotypic features, achieves an average location accuracy reaching 82%, significantly improving the precision of harvesting point localization for lychee clusters.

Cucumber Picking Recognition in Near-Color Background Based on Improved YOLOv5

Rapid and precise detection of cucumbers is a key element in enhancing the capability of intelligent harvesting robots. Problems such as near-color background interference, branch and leaf occlusion of fruits, and target scale diversity in greenhouse environments posed higher requirements for cucumber target detection algorithms. Therefore, a lightweight YOLOv5s-Super model was proposed based on the YOLOv5s model. First, in this study, the bidirectional feature pyramid network (BiFPN) and C3CA module were added to the YOLOv5s-Super model with the goal of capturing cucumber shoulder features of long-distance dependence and dynamically fusing multi-scale features in the near-color background. Second, the Ghost module was added to the YOLOv5s-Super model to speed up the inference time and floating-point computation speed of the model. Finally, this study visualized different feature fusion methods for the BiFPN module; independently designed a C3SimAM module for comparison between parametric and non-parametric attention mechanisms. The results showed that the YOLOv5s-Super model achieves mAP of 87.5%, which was 4.2% higher than the YOLOv7-tiny and 1.9% higher than the YOLOv8s model. The improved model could more accurately and robustly complete the detection of multi-scale features in complex near-color backgrounds while the model met the requirement of being lightweight. These results could provide technical support for the implementation of intelligent cucumber picking.

APPLICATION OF MACHINE LEARNING IN ESTIMATING ONTREE YIELD OF CITRUS FRUIT

Citrus is manually counted to estimate the yield. By using some innovative agricultural techniques yield and production can be increased. Numerous agricultural innovations have been introduced in recent years. Higher agricultural production, prediction, and reliable crop status information are more important than ever due to the expected growth of the human population. Agriculture has always been the foundation of human society. Current study was aimed to develop a reliable and meaningful information-gathering agricultural field based on image processing during 2020. Citrus yield can be increased in the initial stages by counting it with RGB and HSV-based images taken from an Android phone from various angles using machine learning techniques. Fertilizers such as potash, phosphorus, and nitrogen can then be utilized to boost yield. According to the findings, farmers can control and monitor citrus health production more efficiently and effectively by integrating machine learning with agriculture. The citrus calculation using the given technique compared with manually counted citrus, having difference of up to 5 to 10 citruses for a single plant per plot in a field. The proposed method produced excellent results under varying lighting conditions, leaf occlusion, and fruit overlap on photos taken at various distances from the orange trees.

A Single-Tree Point Cloud Completion Approach of Feature Fusion for Agricultural Robots

With the continuous development of digital agriculture and intelligent forestry, the demand for three-dimensional modeling of trees or plants using agricultural robots is also increasing. Laser radar technology has gradually become an important technical means for agricultural robots to obtain three-dimensional information about trees. When using laser radar to scan trees, incomplete point cloud data are often obtained due to leaf occlusion, visual angle limitation, or operation error, which leads to quality degradation of the subsequent 3D modeling and quantitative analysis of trees. At present, a lot of research work has been carried out in the direction of point cloud completion, in which the deep learning model is the mainstream solution. However, the existing deep learning models have mainly been applied to urban scene completion or the point cloud completion of indoor regularized objects, and the research objects generally have obvious continuity and symmetry characteristics. There has been no relevant research on the point cloud completion method for objects with obvious individual morphological differences, such as trees. Therefore, this paper proposes a single-tree point cloud completion method based on feature fusion. This method uses PointNet, based on point structure, to extract the global features of trees, and EdgeConv, based on graph structure, to extract the local features of trees. After integrating global and local features, FoldingNet is used to realize the generation of a complete point cloud. Compared to other deep learning methods on the open source data set, the CD index using this method increased by 21.772% on average, and the EMD index increased by 15.672% on average, which proves the effectiveness of the method in this paper and provides a new solution for agricultural robots to obtain three-dimensional information about trees.

An Automatic Jujube Fruit Detection and Ripeness Inspection Method in the Natural Environment

The ripeness phases of jujube fruits are one factor mitigating against fruit detection, in addition to uneven environmental conditions such as illumination variation, leaf occlusion, overlapping fruits, colors or brightness, similar plant appearance to the background, and so on. Therefore, a method called YOLO-Jujube was proposed to solve these problems. With the incorporation of the networks of Stem, RCC, Maxpool, CBS, SPPF, C3, PANet, and CIoU loss, YOLO-Jujube was able to detect jujube fruit automatically for ripeness inspection. Having recorded params of 5.2 m, GFLOPs of 11.7, AP of 88.8%, and a speed of 245 fps for detection performance, including the sorting and counting process combined, YOLO-Jujube outperformed the network of YOLOv3-tiny, YOLOv4-tiny, YOLOv5s, and YOLOv7-tiny. YOLO-Jujube is robust and applicable to meet the goal of a computer vision-based understanding of images and videos.

Research on Instance Segmentation Algorithm of Greenhouse Sweet Pepper Detection Based on Improved Mask RCNN

The fruit quality and yield of sweet peppers can be effectively improved by accurately and efficiently controlling the growth conditions and taking timely corresponding measures to manage the planting process dynamically. The use of deep-learning-based image recognition technology to segment sweet pepper instances accurately is an important means of achieving the above goals. However, the accuracy of the existing instance segmentation algorithms is seriously affected by complex scenes such as changes in ambient light and shade, similarity between the pepper color and background, overlap, and leaf occlusion. Therefore, this paper proposes an instance segmentation algorithm that integrates the Swin Transformer attention mechanism into the backbone network of a Mask region-based convolutional neural network (Mask RCNN) to enhance the feature extraction ability of the algorithm. In addition, UNet3+ is used to improve the mask head and segmentation quality of the mask. The experimental results show that the proposed algorithm can effectively segment different categories of sweet peppers under conditions of extreme light, sweet pepper overlap, and leaf occlusion. The detection AP, AR, segmentation AP, and F1 score were 98.1%, 99.4%, 94.8%, and 98.8%, respectively. The average FPS value was 5, which can be satisfied with the requirement of dynamic monitoring of the growth status of sweet peppers. These findings provide important theoretical support for the intelligent management of greenhouse crops.

Point Cloud Completion of Plant Leaves under Occlusion Conditions Based on Deep Learning.

The utilization of 3-dimensional point cloud technology for non-invasive measurement of plant phenotypic parameters can furnish important data for plant breeding, agricultural production, and diverse research applications. Nevertheless, the utilization of depth sensors and other tools for capturing plant point clouds often results in missing and incomplete data due to the limitations of 2.5D imaging features and leaf occlusion. This drawback obstructed the accurate extraction of phenotypic parameters. Hence, this study presented a solution for incomplete flowering Chinese Cabbage point clouds using Point Fractal Network-based techniques. The study performed experiments on flowering Chinese Cabbage by constructing a point cloud dataset of their leaves and training the network. The findings demonstrated that our network is stable and robust, as it can effectively complete diverse leaf point cloud morphologies, missing ratios, and multi-missing scenarios. A novel framework is presented for 3D plant reconstruction using a single-view RGB-D (Red, Green, Blue and Depth) image. This method leveraged deep learning to complete localized incomplete leaf point clouds acquired by RGB-D cameras under occlusion conditions. Additionally, the extracted leaf area parameters, based on triangular mesh, were compared with the measured values. The outcomes revealed that prior to the point cloud completion, the R2 value of the flowering Chinese Cabbage's estimated leaf area (in comparison to the standard reference value) was 0.9162. The root mean square error (RMSE) was 15.88 cm2, and the average relative error was 22.11%. However, post-completion, the estimated value of leaf area witnessed a significant improvement, with an R2 of 0.9637, an RMSE of 6.79 cm2, and average relative error of 8.82%. The accuracy of estimating the phenotypic parameters has been enhanced significantly, enabling efficient retrieval of such parameters. This development offers a fresh perspective for non-destructive identification of plant phenotypes.

Recognition and Location of Pepper Picking Based on Improved YOLOv5s and Depth Camera

HighlightsAn improved YOLOv5s deep learning model was used to identify peppers in complex background.The deep-level features on 3D (O-XYZ) coordinate of peppers were extracted using RealSense depth camera.An image database set of pepper in different scenes was established.A pepper recognition and location system were constructed based on improved YOLOv5s network.The proposed method achieved a mean average precision of 95.6% and minimum depth error of 0.001 m.Abstract. In order to investigate the impact of different scenes on the recognition performance and obtain the location information of picking targets, the recognition and location system based on improved YOLOv5s network and RealSense depth camera was constructed in this study. An image database in different scenes was established including light intensity, occlusion and overlap degree of pepper. An improved YOLOv5s deep learning model with bidirectional feature pyramid network (BiFPN) was used for the deep feature extraction and high-precision detection of pepper, and the effects of different scenes on recognition accuracy of the model were studied. The results showed that mean average precision (mAP) of YOLOv5s model reached 0.956, which was respectively 6.1%, 9.3%, 44.4%, and 8.2% higher than that of YOLOv4, YOLOv3, YOLOv2, and Faster R-CNN model. The model had good robustness under daytime and evening scenes with the mAP value higher than 0.9. The detection accuracy of the model in the leaf occlusion scenes was better than that of fruit overlap. The detection error was 0.001m which could not affect the picking positioning precision when the Z value of three-dimensional coordinates (O-XYZ) of pepper was 0.2 m. The improved algorithm can accurately recognize and extract three-dimensional coordinates of pepper, which reduces the calculations by eliminating lots of duplicate and redundant prediction boxes and provides a reference for trajectory planning of pepper picking operation. Keywords: Different scenes, Pepper recognition and location, Picking operation, YOLOv5s.

A Real-Time Detection Algorithm for Sweet Cherry Fruit Maturity Based on YOLOX in the Natural Environment

Fast, accurate, and non-destructive large-scale detection of sweet cherry ripeness is the key to determining the optimal harvesting period and accurate grading by ripeness. Due to the complexity and variability of the orchard environment and the multi-scale, obscured, and even overlapping fruit, there are still problems of low detection accuracy even using the mainstream algorithm YOLOX in the absence of a large amount of tagging data. In this paper, we proposed an improved YOLOX target detection algorithm to quickly and accurately detect sweet cherry ripeness categories in complex environments. Firstly, we took a total of 2400 high-resolution images of immature, semi-ripe, and ripe sweet cherries in an orchard in Hanyuan County, Sichuan Province, including complex environments such as sunny days, cloudy days, branch and leaf shading, fruit overlapping, distant views, and similar colors of green fruits and leaves, and formed a dataset dedicated to sweet cherry ripeness detection by manually labeling 36068 samples, named SweetCherry. On this basis, an improved YOLOX target detection algorithm YOLOX-EIoU-CBAM was proposed, which embedded the Convolutional Block Attention Module (CBAM) between the backbone and neck of the YOLOX model to improve the model’s attention to different channels, spaces capability, and replaced the original bounding box loss function of the YOLOX model with Efficient IoU (EIoU) loss to make the regression of the prediction box more accurate. Finally, we validated the feasibility and reliability of the YOLOX-EIoU-CBAM network on the SweetCherry dataset. The experimental results showed that the method in this paper significantly outperforms the traditional Faster R-CNN and SSD300 algorithms in terms of mean Average Precision (mAP), recall, model size, and single-image inference time. Compared with the YOLOX model, the mAP of this method is improved by 4.12%, recall is improved by 4.6%, F-score is improved by 2.34%, while model size and single-image inference time remain basically comparable. The method in this paper can cope well with complex backgrounds such as fruit overlap, branch and leaf occlusion, and can provide a data base and technical reference for other similar target detection problems.

Fruit detection and positioning technology for a Camellia oleifera C. Abel orchard based on improved YOLOv4-tiny model and binocular stereo vision

In the complex environment of an orchard, changes in illumination, leaf occlusion, and fruit overlap make it challenging for mobile picking robots to detect and locate oil-seed camellia fruit. To address this problem, YOLO-Oleifera was developed as a fruit detection model method based on a YOLOv4-tiny model, To obtain clustering results appropriate to the size of the Camellia oleifera fruit, the k-means++ clustering algorithm was used instead of the k-means clustering algorithm used by the YOLOv4-tiny model to determine bounding box priors. Two convolutional kernels of 1 × 1 and 3 × 3 were respectively added after the second and third CSPBlock modules of the YOLOv4-tiny model. This model allows the learning of Camellia oleifera fruit feature information and reduces overall computational complexity. Compared with the classic stereo matching method based on binocular camera images, this method innovatively used the bounding box generated by the YOLO-Oleifera model to extract the region of interest of the fruit, and then adaptively performs stereo matching according to the generation mechanism of the bounding box. This allows the determination of disparity and facilitates the subsequent use of the triangulation principle to determine the picking position of the fruit. An ablation experiment demonstrated the effective improvement of the YOLOv4-tiny model. Camellia oleifera fruit images obtained under sunlight and shading conditions were used to test the YOLO-Oleifera model, and the model robustly detected the fruit under different illumination conditions. Occluded Camellia oleifera fruit decreased precision and recall due to the loss of semantic information. Comparison of this model with deep learning models YOLOv5-s,YOLOv3-tiny, and YOLOv4-tiny, the YOLO-Oleifera model achieved the highest AP of 0.9207 with the smallest data weight of 29 MB. The YOLO-Oleifera model took an average of 31 ms to detect each fruit image, fast enough to meet the demand for real-time detection. The algorithm exhibited high positioning stability and robust function despite changes in illumination. The results of this study can provide a technical reference for the robust detection and positioning of Camellia oleifera fruit by a mobile picking robot in a complex orchard environment.